I Smelled Something Fishy: The Napkin Math That Predicted the SpaceX IPO Windfall

Kyle Hill was CERTAIN SpaceX is Vaporware

When Kyle Hill’s video “Space Data Centers Are Dumb” dropped, the internet went wild. Suddenly, the “smart consensus” was that putting AI compute into orbit was an engineering joke. The video claimed it would take 350 ISS-sized solar arrays just to power a single data center in space.

On its face, the tech community ate it up. “Space is a vacuum! You can’t cool a data center without air! The solar requirements are impossible!”

I’ll admit, the thermal physics are brutal—you can’t use fans to cool silicon when there’s no air to blow. But the moment the video claimed the solar power scaling was impossible, my terrestrial solar spidey sense tingled. Something was fishy. The critics were treating orbital compute like a monolithic 1970s Skylab project. They assumed some idiot was trying to launch a single, centralized megastructure into orbit.

I sat down, did some quick napkin math on distributed satellite networks, and had an easy epiphany: The critics were off the reservation, and SpaceX was about to absolutely rock the world at its upcoming IPO. I usually roll my eyes when people say someone is “playing 4D chess,” but I thought xAI was the underdog. But folding into SpaceX, they’ve flipped the table.

As we now know, after the historic largest public listing in history, many investors agree with me. Here is the napkin math that got me there.

Starlink vs. ISS Reality Check

The core flaw of Kyle Hill’s “350 ISS arrays” argument is that it treats the International Space Station as the pinnacle of solar deployment. It isn’t. It’s an old fart TBH. Just because it’s the single biggest doesn’t mean it’s the best. Distributed modular networks are already changing the game.

Let’s look at the actual solar footprint currently operating in orbit:

• The ISS Baseline: The International Space Station has a total solar surface area of roughly 2,500 square meters (~27,000 sq ft), generating about 120kW to 240kW depending on alignment.
• The Single Starlink Footprint: A standard Starlink V2 Mini satellite features a massive dual-array setup totaling roughly 105 square meters of solar panels per bird.
• The Constellation Scale: SpaceX currently has over 6,000 active satellites in orbit. Even using a conservative average across older and newer models, the active Starlink solar footprint is sitting at roughly 500,000 square meters.
• The Ratio: 500,000 m² (Starlink) ÷ 2,500 m² (ISS) = 200×

The active Starlink communication network already has 200X more solar panel surface area deployed in space than the ISS. SpaceX didn’t build one massive array; they distributed it across a mesh network. If they could do that for internet transit, why not for compute?

Phase 2: Enter the “Starmind” Data Center Satellites

The xAI merger ahead of the listing made the vision crystal clear. SpaceX isn’t planning on running ChatGPT on a standard internet satellite. They are building Starmind, a dedicated constellation of up to one million orbital compute nodes.

When Elon Musk unboxed the engineering specifications for the prototype AI1 Compute Satellite, the numbers completely validated my thesis:

• The Power Profile: Each individual AI1 satellite is designed to pull a peak of 150 kW of solar power to support roughly 120 kW of pure, average AI compute.
• The Wing Scale: To capture this, the AI1 features a massive 70-meter solar wing paired with specialized double-sided radiators to solve that “brutal” vacuum cooling problem.
• The Single-Node Generation: At 150 kW per satellite, a single Starmind node generates more than half the peak power capacity of the entire ISS, packaged into a single launchable unit.

The Math on a 10,000-Node Cluster:

If SpaceX deploys a modest initial cluster of just 10,000 Starmind AI satellites—a fraction of their approved Starlink shell footprint—the numbers become staggering:

10,000 satellites × 150 kW = 1.5 Gigawatts

A 1.5-gigawatt distributed data center in orbit completely bypasses terrestrial power grids, land rights, local server water bills, and community pushback.

The Starship Multiplier That Won Wall Street

The final piece of the puzzle that the “space data centers are dumb” crowd ignored was the launch vehicle logistics. They calculated launch costs using legacy rocketry metrics.

But with Starship V3 moving into operational deployment, SpaceX can lift over 100+ metric tons to LEO per single flight. A single Starship flight has the payload capacity of five Falcon 9 launches combined. More importantly, it features the volumetric fairing size required to deploy these massive 70-meter solar-wing AI satellites en masse.

When SpaceX finally went public, the market didn’t just price them as a launch provider or an ISP. The market priced them as the future backbone of global AI infrastructure.

While the skeptics were busy laughing at theoretical GPU limits and counting ISS panels, the napkin math showed that SpaceX was quietly building a trillion-dollar energy and compute moat.

The Final Nail in the Coffin: Every Big Tech Giant is a Tenant for SpaceX Now

Anthropic is already paying SpaceX $50 million-per-MegaWatt-per-year to rent current ground capacity, and they need more. Ground capacity will only get more expensive from here as grid supply taps out and regulators tighten the noose on behalf of an angry public. Big tech would be willing to pay MORE for space capacity, not less, due to the raw convenience of less beaurocratic red tape. But miraculously, space compute will cost SpaceX LESS, not more.

Based on yet more napkin math RE: cost to deploy enough AI1 satellites to replace Anthropic’s deal with a space-based equivalent, I guesstimate SpaceX has an amortized cost of ~$30 million for EITHER space-based or land-based new compute build-out. This assumes NO new regulation for ground-based data centers other than requirement to generate their own off-grid power. In other words, SpaceX can pay off new Colossus cap-ex in less than 1 year with its current Anthropic deal, but they’ll be able to do approximately the same in space. Meanwhile, the government will seek a much larger pound of flesh from the servers on the ground, making space the cheaper for SpaceX while the supply can’t catch up to demand of all other competitors stuck on the ground.

Amazon, Google, Microsoft, Meta, Oracle, OpenAI, and SoftBank are all building hundreds of ground data centers. Red tape will eat them alive, creating intense development lag, leading to shortfalls in available ground compute versus projections. They will thus have no choice but to rent from SpaceX to close the gap. It will be the only available option. Space compute will cost SpaceX less, but because of the demand squeeze, they will charge a premium for it.

Checkmate.

Kyle Hill walks back the original doomer claims in his first video